Best regards *Timo Wischer* Engineering Software Base (ADITG/ESB)
Tel. +49 5121 49 6938 On 10/18/18 16:58, Takashi Iwai wrote:
On Thu, 18 Oct 2018 16:31:21 +0200, Timo Wischer wrote:
On 10/18/18 15:19, Takashi Iwai wrote:
On Thu, 18 Oct 2018 14:28:47 +0200, Timo Wischer wrote:
On 10/18/18 14:02, Takashi Iwai wrote:
On Thu, 18 Oct 2018 13:33:24 +0200, twischer@de.adit-jv.com wrote:
From: Timo Wischer twischer@de.adit-jv.com
Without this change an interval of (x x+1] will be interpreted as an empty interval but the right value would be x+1. This leads to a failing snd_pcm_hw_params() call which returns -EINVAL.
An example issue log is given in the following: snd_pcm_hw_params failed with err -22 (Invalid argument) ACCESS: MMAP_NONINTERLEAVED FORMAT: S16_LE SUBFORMAT: STD SAMPLE_BITS: 16 FRAME_BITS: 16 CHANNELS: 1 RATE: 16000 PERIOD_TIME: (15999 16000] PERIOD_SIZE: (255 256] PERIOD_BYTES: (510 512] PERIODS: [2 3) BUFFER_TIME: 32000 BUFFER_SIZE: 512 BUFFER_BYTES: 1024
In case of (x x+1) we have to interpret it anyway as a single value of x to compensate rounding issues. For example the period size will result in an interval of (352 353) when the period time is 16ms and the sample rate 22050 Hz (16ms * 22,05 kHz = 352,8 frames). But 352 has to be chosen to allow a buffer size of 705 (32ms * 22,05 kHz = 705,6 frames) which has to be >= 2x period size to avoid Xruns. The buffer size will not end up with an interval of (705 706) simular to the period size because snd_pcm_rate_hw_refine_cchange() calls snd_interval_floor() for the buffer size. Therefore this value will be interpreted as an integer interval instead of a real interval further on.
This issue seems to exist since the change of 9bb985c38 ("pcm: snd_interval_refine_first/last: exclude value only if also excluded before")
Signed-off-by: Timo Wischer twischer@de.adit-jv.com
> On 10/18/18 12:57, Takashi Iwai wrote: > This change looks risky. The snd_interval_value() might be called > even if the interval isn't reduced to a single value. Rather check > openmin instead. If I would do "if (i->openmin)" on an interval of (x x+1) x+1 would be returned. But this would result in the second issue which I have tried to describe in the commit message: x has to be returned otherwise it is not guaranteed that buffer_size >= 2x period_size. And this would result in continuously Xruns. Therefore I would like to prefer "if (i->openmin && !i->openmax)"
Hm, I overlooked that there is an assert() before that. So a single-value interval is a must at this execution, so it doesn't matter which one to take.
Didn't the assert() hit in your case with x+1, then?
No. For the implementation of
snd_interval_value() (x x+1) is also a valid interval which has only a single value. This is also required because the ALSA neogation rules often end up with intervals like (x x+1) e.g. in case of rounding issues 32ms * 22,05 kHz = 705,6 frames -> (705 706) Therefore assert(snd_interval_single(i)) has to not fail in case of (705 706).
And this is a problematic case no matter which value to take. Your change might "fix" your specific case, but it might break others. No matter which value (min or max) to take, it's basically invalid.
If your fix is logically correct, it's fine. But, judging from the description, it looks like nothing but heuristic based on your specific use cases.
The reason for this fix was an issue where an interval of (x x+1] is interpreted as not a single value. Before the change in snd_interval_single() the following intervals were interpreted as a single value:
- (x x) -> required for rounding issues e.g. 32ms * 22,05 kHz = 705,6 frames => (705 706)
Wait... Isn't it (x x+1) at this point? Otherwise they can't be a valid value, per definition.
Yes, you are right. May mistake. I have also no idea why and if these intervals required
* (x x) *
[x x)
*
[x x]
At least snd_interval_checkempty() is interpreting them correctly and returning true
- (x x]
- [x x)
- [x x]
- (x x+1)
- [x x+1)
With this change (x x+1] will also be interpreted as a single value (which looks reasonable for me). But the interval [x x+1] is still not interpreted as a single value (which also sounds right for me)
Before the change in snd_interval_value() for all intervals x was returned. With this change only in case (x x+1] x+1 will be returned. All other intervals are not changed.
Therefore this is a minimal change which solves issues. I have already at least two use cases which are failing without this fix. (Somehow simple use cases with rate->dmix->hw where the hw truncates the default period_time of 125ms and fails with EINVAL on snd_pcm_set_period_near()). Therefore I would expect that these issues will be seen soon by other users.
If there is a use case which breaks in future due to the right interpretation of (x x+1] I think we should try to find the root cause why this (x x+1] interval has to be interpreted wrongly to get the use case to work.
I'm not against fixing the handling of (x x+1] or (x x+1). Although (x x+1) can't be handled correctly at obtaining the exact value, often PERIOD_TIME or BUFFER_TIME may reach to this state, so allowing it still makes sense.
But what I don't understand is why
snd_interval_value() { if (openmin) return min+1; return min; }
doesn't work. I don't mean that this is better than returning max (like your v1 patch), but just want to understand the problem correctly.
In the issue case which I had I got a buffer size < 2x period size while I have requested 16ms period and 32ms buffer. Unfortunately this results in continuously under runs. The following audio path was used: aplay->rate->dmix->hw. aplay streams with 22050 Hz. Therefore we got the following intervals: period: 16ms * 22,05 kHz = 352,8 frames => (352 353) buffer: 32ms * 22,05 kHz = 705,6 frames => [705 706] The buffer size will not end up with an interval of (705 706) similar to the period size because snd_pcm_rate_hw_refine_cchange() calls snd_interval_floor() for the buffer size. Therefore this value will be interpreted as an integer interval (integer flag is set) instead of a real interval further on.
With your solution we would use 353 frames for period and 705 frames for buffer. Therefore we will get Xruns because 2*353 > 705.
IOW, if we have a good set of unit tests to cover most of possible use cases and this is proven to improve all these, I'll happily take it. But unfortunately there is no good test coverage, so far.
At least I have executed several tests on my end which are using different combination of ALSA plugins (plug, rate, dmix, dsnoop), sample rates, channels, period and buffer sizes on different hardware with different sound cards. I would expect that you are talking about the tests in the test directory of the alsa-lib sources. Or is there anything else? May be I could try to extend/add some tests at least to reproduce the issue of this patch.
I'd think of a test program linked with these functions for executing only this configuration stuff without actual plugins. Then we can give a bunch of sets to evaluate the extreme cases.
So you mean only testing the corner cases of the functions of src/pcm/interval_inline.h without any other alsa-lib dependencies?
thanks,
Takashi